TimesFM

TimesFM (Time Series Foundation Model) is a pretrained decoder-only transformer built specifically for time-series forecasting, developed by Google Research. Unlike general LLMs adapted to forecasting, it is trained natively on temporal data and produces forecasts zero-shot, without task-specific fine-tuning.

The model evolved quickly: TimesFM 2.5 (200M parameters, 2025) became a production model inside BigQuery ML's AI.FORECAST, and TimesFM 3.0 (330M parameters, September 2026) added native multivariate forecasting with covariate support, ranking first across the GIFT-Eval, FEV-Bench, and TIME leaderboards. Licensing nuance: code and weights through 2.5 are Apache 2.0, but TimesFM 3.0 weights ship under a non-commercial license, so check the license before production use.